## AI and Scientific Research: Towards New Frontiers of Creativity The integration of Large Language Models (LLMs) into the scientific ecosystem raises fundamental questions about the ability of AI to generate truly new and original ideas. A recent study focused precisely on this aspect, analyzing whether complex workflows, employing iterative reasoning, evolutionary algorithms, and recursive decomposition, can lead to more innovative and feasible research plans. ## Agentic Workflows: A Comparison of Architectures The research compared five different reasoning architectures: * Reflection-based iterative refinement. * Sakana AI v2 evolutionary algorithms. * Google Co-Scientist multi-agent framework. * GPT Deep Research (GPT-5.1) recursive decomposition. * Gemini 3 Pro multimodal long-context pipeline. Thirty proposals were evaluated for each architecture, considering originality, feasibility, and potential impact. The results showed that workflows based on decomposition and long-context analysis achieve a mean novelty score of 4.17/5, while reflection-based approaches score significantly lower (2.33/5). ## Implications and Future Perspectives The study found variability in performance depending on the research domain, with the highest-performing workflows able to maintain feasibility without sacrificing creativity. These findings suggest that carefully designed multi-stage agentic workflows can effectively improve the ideation process in AI-assisted scientific research. This opens new perspectives for the use of AI as a tool to support creativity and innovation in the scientific field.